Research On Compressed Sensing And Reconstruction Of Electrical Equipment Images Under Deep Learning Framework

To ensure the stable running of the power transmission system,various electrical equipment in the power system must be monitored at all time.The collection,transmission and reconstruction of electrical equipment images play an important role in the real-time monitoring of equipment.However,the sampling of a large number of electrical equipment images will cause huge storage and transmission burden.How to effectively reduce the sampling rate of electrical equipment images and reconstruct images with high quality is a problem that needs to be solved urgently.For these problems,this thesis proposes compressed sensing and reconstruction algorithm of electrical equipment images under the deep learning framework,validates and analyzes the proposed algorithm through experiments.The main work of this thesis is as follows:(1)In terms of the problem of current based on deep learning image compressed sensing reconstruction algorithms with poor reconstruction quality and obvious blocking effect,this thesis proposes a multi-scale convolutional network(Ms RFCNN-3)for electrical equipment image compressed sensing reconstruction.In the image sampling part,this paper uses convolutional neural network sampling to replace the traditional block measurement matrix sampling,which avoids the blocking effect caused by the block processing of the image before sampling and provides more structural information for subsequent image reconstruction.In the image reconstruction part,this paper designs a multi-scale feature extraction module to extract the multi-scale feature information in the image,and the dilated convolution kernel is used in this module to avoid the increasing of parameters caused by the large-size convolution kernel.Compared with the single-channel convolutional network,the reconstruction accuracy of this algorithm is improved by 0.49 d B.(2)Aiming at the problem that the feature information extracted by the shallow neural network is limited,and each feature map in the network is treated equally,which hinders the network’s representation ability,this thesis proposes a multi-scale residual network(AD-Ms RFCNN)for electrical equipment image compressed sensing reconstruction.On the basis of Ms RFCNN-3,the residual network is applied to increase the network depth,and the short skip in the residual block can reduce the loss of information and obtains richer image information,which protects the integrity of the data information,and attention mechanism is added to the network to adaptively scale each feature map to obtain more detailed information that the target needs to pay attention to.Under the same test set,the PSNR value of this algorithm is 0.39 d B higher than that of Ms RFCNN-3.(3)In view of the lack of feedback mechanism in the current compressed sensing reconstruction network,which restricts the improvement of reconstruction quality,this thesis proposes multi-scale generative confrontation network(Ms GAN)framework to realize compressed sensing reconstruction of electrical equipment images.The algorithms use Ms RFCNN-3 and AD-Ms RFCNN as the generation network respectively,and discriminant network that has a feedback on the image generation process is added.The discriminant network is trained to distinguish the reconstructed image and the original real image,so that the reconstruction result is closer to the original real image.Experimental results show that after adds the adversarial network,the reconstruction quality is improved by 0.27 d B compared with the Ms RFCNN-3 algorithm and is improved by 0.55 d B compared with the AD-Ms RFCNN algorithm.